From d8e8c7dda11f5d5cf90495f2e89d917a83509bc0 Mon Sep 17 00:00:00 2001 From: Vineet Gupta Date: Thu, 28 Nov 2013 13:57:54 +0530 Subject: ARC: [SMP] optimize IPI send and receive * Don't send an IPI if receiver already has a pending IPI. Atomically piggyback the new msg with pending msg. * IPI receiver looping on xchg() not required References: https://lkml.org/lkml/2013/11/25/232 Suggested-by: Peter Zijlstra Signed-off-by: Vineet Gupta --- arch/arc/kernel/smp.c | 68 ++++++++++++++++++++++++++++++--------------------- 1 file changed, 40 insertions(+), 28 deletions(-) (limited to 'arch/arc') diff --git a/arch/arc/kernel/smp.c b/arch/arc/kernel/smp.c index c00c612..40859e5 100644 --- a/arch/arc/kernel/smp.c +++ b/arch/arc/kernel/smp.c @@ -215,16 +215,31 @@ static DEFINE_PER_CPU(unsigned long, ipi_data); static void ipi_send_msg_one(int cpu, enum ipi_msg_type msg) { unsigned long __percpu *ipi_data_ptr = per_cpu_ptr(&ipi_data, cpu); + unsigned long old, new; unsigned long flags; pr_debug("%d Sending msg [%d] to %d\n", smp_processor_id(), msg, cpu); local_irq_save(flags); - set_bit(msg, ipi_data_ptr); + /* + * Atomically write new msg bit (in case others are writing too), + * and read back old value + */ + do { + new = old = *ipi_data_ptr; + new |= 1U << msg; + } while (cmpxchg(ipi_data_ptr, old, new) != old); - /* Call the platform specific cross-CPU call function */ - if (plat_smp_ops.ipi_send) + /* + * Call the platform specific IPI kick function, but avoid if possible: + * Only do so if there's no pending msg from other concurrent sender(s). + * Otherwise, recevier will see this msg as well when it takes the + * IPI corresponding to that msg. This is true, even if it is already in + * IPI handler, because !@old means it has not yet dequeued the msg(s) + * so @new msg can be a free-loader + */ + if (plat_smp_ops.ipi_send && !old) plat_smp_ops.ipi_send(cpu); local_irq_restore(flags); @@ -269,31 +284,23 @@ static void ipi_cpu_stop(void) machine_halt(); } -static inline void __do_IPI(unsigned long pending) +static inline void __do_IPI(unsigned long msg) { - while (pending) { - - unsigned long msg = __ffs(pending); - - switch (msg) { - case IPI_RESCHEDULE: - scheduler_ipi(); - break; + switch (msg) { + case IPI_RESCHEDULE: + scheduler_ipi(); + break; - case IPI_CALL_FUNC: - generic_smp_call_function_interrupt(); - break; - - case IPI_CPU_STOP: - ipi_cpu_stop(); - break; - - default: - pr_warn("IPI missing msg\n"); + case IPI_CALL_FUNC: + generic_smp_call_function_interrupt(); + break; - } + case IPI_CPU_STOP: + ipi_cpu_stop(); + break; - pending &= ~(1U << msg); + default: + pr_warn("IPI with unexpected msg %ld\n", msg); } } @@ -312,11 +319,16 @@ irqreturn_t do_IPI(int irq, void *dev_id) plat_smp_ops.ipi_clear(irq); /* - * XXX: is this loop really needed - * And do we need to move ipi_clean inside + * "dequeue" the msg corresponding to this IPI (and possibly other + * piggybacked msg from elided IPIs: see ipi_send_msg_one() above) */ - while ((pending = xchg(this_cpu_ptr(&ipi_data), 0)) != 0) - __do_IPI(pending); + pending = xchg(this_cpu_ptr(&ipi_data), 0); + + do { + unsigned long msg = __ffs(pending); + __do_IPI(msg); + pending &= ~(1U << msg); + } while (pending); return IRQ_HANDLED; } -- cgit v1.1